Bayesian Anytime Pareto Set Identification for Multi-Objective Multi-Armed Bandits

Researchers have introduced two new algorithms to improve decision-making in complex environments: Top-Two Pareto Front Thompson Sampling (TTPFTS) for multi-objective decision-making and Flickering Multi-Armed Bandits (FMAB) for sequential decision-making in changing environments.

TTPFTS, introduced in a paper submitted on June 17, 2026^[1], is the first anytime Multi-Objective Multi-Armed Bandit algorithm for Pareto Set Identification. It takes a Bayesian approach and has been benchmarked against state-of-the-art fixed-budget Pareto Set Identification algorithms. The algorithm was demonstrated to be effective in a challenging multi-objective molecular discovery setting. A novel uncertainty quantification metric was also introduced to estimate the algorithm's confidence in the predicted Pareto set. Meanwhile, a separate paper introduced FMAB to model sequential decision-making in environments with changing action availability^[2]. FMAB restricts access to the next action to a subset dependent on the agent's current choice, imposing a dual challenge of information acquisition and navigation overhead. A two-phase lazy random walk algorithm was proposed for robust exploration, and high-probability sublinear regret bounds were established.